In experiment 1 which investigates hypothesis 5 (i.e. trend timing method “before” is better than trend
timing method “after”), some interesting results are discovered. In addition to supporting hypothesis 5
(only when the refining method are not used), it also returns some surprising differences between the
two methods. The trend timing method “before” is much better at correctly labeling positive
documents than it is as labeling negative documents, and trend timing method “after” is much better at
correctly labeling negative news articles than it is at labeling positive news articles. See section 6.1.2
for experiment analysis.
The main purpose of experiment 2 is to investigate hypothesis 4 (i.e. label refining methods improves
the labels given by the basic price trend labeling method). This experiment clearly shows that label
73
refining helps. In FIGURE 6.5, where the classifier is tested on a automatically labeled test set, the FMeasure
goes from 0.35 with the basic labeling to 0.72 with label refining method two when trend
timing method “before” is used. When trend timing method after is used the F-Measure goes from
0.05 to 0.63. In FIGURE 6.7, where the classifier are tested on the manually labeled data set, the FMeasure
goes from 0.32 with the basic labeling to 0.42 with label refining method two when trend
timing method “before” is used. When trend timing method after is used the F-Measure goes from
0.07 to 0.44. This clearly shows that label refining improves the labels labeled by the basic price trend
method. See section 6.2.2 for experiment analysis.
Experiment 3 simulates trades by using the labels given by the trained classifier. Its main purpose is to
investigate hypothesis 1 and 2 (the system should perform better than random trading and it should
perform on the same level or better as trading with the manually labeled set). Both hypothesizes are
supported by the evidence found in this experiment. The results show that some of the methods
manage to get 0.5% return on average for each trade they perform. Which is promising given that each
trade is only lasting one day. This return is 15% of what the maximum possible return is. See section
6.3.2 for experiment analysis.
The news based trade decision support system proposed in this thesis is shown to be able to do some
good trading decisions. Systems like this can therefore be very beneficial for traders as a tool to help
making better trading decisions. With such a model it is easier to foresee future behaviors and
movement of stock prices. Thus it is also easier to take correct actions immediately and act properly in
trading decisions to gain more profits and prevent losses.